The task of multi-label temporal action detection aims to accurately detect dense action instances in untrimmed videos. Previous methods focused on modeling the appearance features of RGB images have struggled to capture the fine details and subtle variations in human actions, resulting in three critical issues: overlapping action confusion, intra-class appearance diversity, and background interferences. These issues have significantly undermined the accuracy and generalization of detection models. To tackle these issues, we propose incorporating the human skeleton into the feature design of the detection model. By utilizing multi-person skeletons, our proposed method can accurately represent various human actions in the scene, balance the salience of overlapping actions, and reduce the impact of changes in human appearance and background interferences on action features. Overall, we propose a novel two-stream human action aware network~(HAAN) for multi-label temporal action detection based on the original RGB frames and the estimated skeleton frames. To leverage the complementary advantages of RGB features and skeleton features, we design a cross-modality fusion module that allows the two features to guide each other and enhance their representation of human actions. On the popular benchmarks MultiTHUMOS and Charades, our HAAN achieves state-of-the-art performance with 56.9% (+5.4%) and 32.1% (+3.3%) mean average precision (mAP) compared to the best available methods. Importantly, HAAN shows superior improvements of +6.83%, +22.35%, and +2.56% on the challenging sample subsets of the three critical issues.
论文检索
输入标题、作者或关键词，从 1,014 篇学术成果中精准定位
Face analysis tasks have a wide range of applications, but the universal facial representation has only been explored in a few works. In this paper, we explore high-performance pre-training methods to boost the face analysis tasks such as face alignment and face parsing. We propose a self-supervised pre-training framework, called Mask Contrastive Face (MCF), with mask image modeling and a contrastive strategy specially adjusted for face domain tasks. To improve the facial representation quality, we use feature map of a pre-trained visual backbone as a supervision item and use a partially pre-trained decoder for mask image modeling. To handle the face identity during the pre-training stage, we further use random masks to build contrastive learning pairs. We conduct the pre-training on the LAION-FACE-cropped dataset, a variants of LAION-FACE 20M, which contains more than 20 million face images from Internet websites. For efficiency pre-training, we explore our framework pre-training performance on a small part of LAION-FACE-cropped and verify the superiority with different pre-training settings. Our model pre-trained with the full pre-training dataset outperforms the state-of-the-art methods on multiple downstream tasks. Our model achieves 0.932 NME_diag for AFLW-19 face alignment and 93.96 F1 score for LaPa face parsing. Code is available at https://github.com/nomewang/MCF.
Beyond Generic: Enhancing Image Captioning with Real-World Knowledge using Vision-Language Pre-Training Model
PDF ↗Current captioning approaches tend to generate correct but "generic" descriptions that lack real-world knowledge, e.g., named entities and contextual information. Considering that Vision-Language Pre-Training (VLP) models master massive such knowledge from large-scale web-harvested data, it is promising to utilize the generalizability of VLP models to incorporate knowledge into image descriptions. However, using VLP models faces challenges: zero-shot inference suffers from knowledge hallucination that leads to low-quality descriptions, but the generic bias in downstream task fine-tuning hinders the VLP model from expressing knowledge. To address these concerns, we propose a simple yet effective method called Knowledge-guided Replay (K-Replay), which enables the retention of pre-training knowledge during fine-tuning. Our approach consists of two parts: (1) a knowledge prediction task on automatically collected replay exemplars to continuously awaken the VLP model's memory about knowledge, thus preventing the model from collapsing into the generic pattern; (2) a knowledge distillation constraint to improve the faithfulness of generated descriptions hence alleviating the knowledge hallucination. To evaluate knowledge-enhanced descriptions, we construct a novel captioning benchmark KnowCap, containing knowledge of landmarks, famous brands, special foods and movie characters. Experimental results show that our approach effectively incorporates knowledge into descriptions, outperforming strong VLP baseline by 20.9 points (78.7 → 99.6) in CIDEr score and 20.5 percentage points (34.0% → 54.5%) in knowledge recognition accuracy. Our code and data is available at https://github.com/njucckevin/KnowCap.
Uniformly Distributed Category Prototype-Guided Vision-Language Framework for Long-Tail Recognition
PDF ↗Recently, large-scale pre-trained vision-language models have presented benefits for alleviating class imbalance in long-tailed recognition. However, the long-tailed data distribution can corrupt the representation space, where the distance between head and tail categories is much larger than the distance between two tail categories. This uneven feature space distribution causes the model to exhibit unclear and inseparable decision boundaries on the uniformly distributed test set, which lowers its performance. To address these challenges, we propose the uniformly category prototype-guided vision-language framework to effectively mitigate feature space bias caused by data imbalance. Especially, we generate a set of category prototypes uniformly distributed on a hypersphere. Category prototype-guided mechanism for image-text matching makes the features of different classes converge to these distinct and uniformly distributed category prototypes, which maintain a uniform distribution in the feature space, and improve class boundaries. Additionally, our proposed irrelevant text filtering and attribute enhancement module allows the model to ignore irrelevant noisy text and focus more on key attribute information, thereby enhancing the robustness of our framework. In the image recognition fine-tuning stage, to address the positive bias problem of the learnable classifier, we design the class feature prototype-guided classifier, which compensates for the performance of tail classes while maintaining the performance of head classes. Our method outperforms previous vision-language methods for long-tailed learning work by a large margin and achieves state-of-the-art performance.
Localizing objects in 3D scenes according to the semantics of a given natural language is a fundamental yet important task in the field of multimedia understanding, which benefits various real-world applications such as robotics and autonomous driving. However, the majority of existing 3D object grounding methods are restricted to a single-sentence input describing an individual object, which cannot comprehend and reason more contextualized descriptions of multiple objects in more practical 3D cases. To this end, we introduce a new challenging task, called 3D Dense Object Grounding (3D DOG), to jointly localize multiple objects described in a more complicated paragraph rather than a single sentence. Instead of naively localizing each sentence-guided object independently, we found that dense objects described in the same paragraph are often semantically related and spatially located in a focused region of the 3D scene. To explore such semantic and spatial relationships of densely referred objects for more accurate localization, we propose a novel Stacked Transformer based framework for 3D DOG, named 3DOGSFormer. Specifically, we first devise a contextual query-driven local transformer decoder to generate initial grounding proposals for each target object. The design of these contextual queries enables the model to capture linguistic semantic relationships of objects in the paragraph in a lightweight manner. Then, we employ a proposal-guided global transformer decoder that exploits the local object features to learn their correlation for further refining initial grounding proposals. In particular, we develop two types of proposal-guided attention layers to encode both explicit and implicit pairwise spatial relations to enhance 3D relation understanding. Extensive experiments on three challenging benchmarks (Nr3D, Sr3D, and ScanRefer) show that our proposed 3DOGSFormer outperforms state-of-the-art 3D single-object grounding methods and their dense-object variants by significant margins.
Surgical triplet recognition aims to recognize surgical activities as triplets (i.e.,), which provides fine-grained information essential for surgical scene understanding. Existing methods for surgical triplet recognition rely on compositional methods that recognize the instrument, verb, and target simultaneously. In contrast, our method, called chain-of-look prompting, casts the problem of surgical triplet recognition as visual prompt generation from large-scale vision-language (VL) models, and explicitly decomposes the task into a series of video reasoning processes. Chain-of-Look prompting is inspired by: (1) the chain-of-thought prompting in natural language processing, which divides a problem into a sequence of intermediate reasoning steps; (2) the inter-dependency between motion and visual appearance in the human vision system. Since surgical activities are conveyed by the actions of physicians, we regard the verbs as the carrier of semantics in surgical endoscopic videos. Additionally, we utilize the BioMed large language model to calibrate the generated visual prompt features for surgical scenarios. Our approach captures the visual reasoning processes underlying surgical activities and achieves better performance compared to the state-of-the-art methods on the largest surgical triplet recognition dataset, CholecT50. The code is available at https://github.com/southnx/CoLSurgical.
A video storyboard is a roadmap for video creation which consists of shot-by-shot images to visualize key plots in a text synopsis. Creating video storyboards, however, remains challenging which not only requires cross-modal association between high-level texts and images but also demands long-term reasoning to make transitions smooth across shots. In this paper, we propose a new task called Text synopsis to Video Storyboard (TeViS) which aims to retrieve an ordered sequence of images as the video storyboard to visualize the text synopsis. We construct a MovieNet-TeViS dataset based on the public MovieNet dataset [17]. It contains 10K text synopses each paired with keyframes manually selected from corresponding movies by considering both relevance and cinematic coherence. To benchmark the task, we present strong CLIP-based baselines and a novel VQ-Trans model. VQ-Trans first encodes text synopsis and images into a joint embedding space and uses vector quantization (VQ) to improve the visual representation. Then, it auto-regressively generates a sequence of visual features for retrieval and ordering. Experimental results demonstrate that VQ-Trans significantly outperforms prior methods and the CLIP-based baselines. Nevertheless, there is still a large gap compared to human performance suggesting room for promising future work. The code and data are available at: https://ruc-aimind.github.io/projects/TeViS/
Adversarial training (AT) is one of the most effective ways for deep neural network models to resist adversarial examples. However, there is still a significant gap between robust training accuracy and testing accuracy. Although recent studies have shown that data augmentation can effectively reduce this gap, most methods heavily rely on generating large amounts of training data without considering which features are beneficial for model robustness, making them inefficient. To address the above issue, we propose a two-stage AT algorithm for image data that adopts different data augmentation strategies during the training process to improve model robustness. In the first stage, we focus on the convergence of the algorithm, which uses structure and texture information to guide AT. In the second stage, we introduce a strategy that randomly fuses the data features to generate diverse adversarial examples for AT. We compare our proposed algorithm with five state-of-the-art algorithms on three models, and the experimental results achieve the best robust accuracy under all evaluation metrics on the CIFAR10 dataset, demonstrating the superiority of our method.
Text-based image captioning is a vital but under-explored task, which aims to describe images by captions containing scene text automatically. Recent studies have made encouraging progress, but they are still suffering from two issues. Firstly, current models cannot capture and generate scene text in non-Latin script languages, which severely limits the objectivity and the information completeness of generated captions. Secondly, current models tend to describe images with monotonous and templated style, which greatly limits the diversity of the generated captions. Although the above-mentioned issues can be alleviated through carefully designed annotations, this process is undoubtedly laborious and time-consuming. To address the above issues, we propose a Zero-shot Framework for Text-based Image Captioning (Zero-TextCap). Concretely, to generate candidate sentences starting from the prompt 'Image of' and iteratively refine them to improve the quality and diversity of captions, we introduce a Hybrid-sampling masked language model (H-MLM). To read multi-lingual scene text and model the relationships between them, we introduce a robust OCR system. To ensure that the captions generated by H-MLM contain scene text and are highly relevant to the image, we propose a CLIP-based generation guidance module to insert OCR tokens and filter candidate sentences. Our Zero-TextCap is capable of generalizing captions containing multi-lingual scene text and boosting the diversity of captions. Sufficient experiments demonstrate the effectiveness of our proposed Zero-TextCap. Our codes are available at https://github.com/Gemhuang79/Zero_TextCap.
In e-commerce, products and micro-videos serve as two primary carriers. Introducing cross-domain retrieval between these carriers can establish associations, thereby leading to the advancement of specific scenarios, such as retrieving products based on micro-videos or recommending relevant videos based on products. However, existing datasets only focus on retrieval within the product domain while neglecting the micro-video domain and often ignore the multi-modal characteristics of the product domain. Additionally, these datasets strictly limit their data scale through content alignment and use a content-based data organization format that hinders the inclusion of user retrieval intentions. To address these limitations, we propose the PKU Real20M dataset, a large-scale e-commerce dataset designed for cross-domain retrieval. We adopt a query-driven approach to efficiently gather over 20 million e-commerce products and micro-videos, including multimodal information. Additionally, we design a three-level entity prompt learning framework to align inter-modality information from coarse to fine. Moreover, we introduce the Query-driven Cross-Domain retrieval framework (QCD), which leverages user queries to facilitate efficient alignment between the product and micro-video domains. Extensive experiments on two downstream tasks validate the effectiveness of our proposed approaches. The dataset and source code are available at https://github.com/PKU-ICST-MIPL/Real20M_ACMMM2023.
Data-Free Knowledge Distillation (DFKD) has shown great potential in creating a compact student model while alleviating the dependency on real training data by synthesizing surrogate data. However, prior arts are seldom discussed under distribution shifts, which may be vulnerable in real-world applications. Recent Vision-Language Foundation Models, e.g., CLIP, have demonstrated remarkable performance in zero-shot out-of-distribution generalization, yet consuming heavy computation resources. In this paper, we discuss the extension of DFKD to Vision-Language Foundation Models without access to the billion-level image-text datasets. The objective is to customize a student model for distribution-agnostic downstream tasks with given category concepts, inheriting the out-of-distribution generalization capability from the pre-trained foundation models. In order to avoid generalization degradation, the primary challenge of this task lies in synthesizing diverse surrogate images driven by text prompts. Since not only category concepts but also style information are encoded in text prompts, we propose three novel Prompt Diversification methods to encourage image synthesis with diverse styles, namely Mix-Prompt, Random-Prompt, and Contrastive-Prompt. Experiments on out-of-distribution generalization datasets demonstrate the effectiveness of the proposed methods, with Contrastive-Prompt performing the best.
Deep Neural Networks suffer significant performance degeneration when noisy labels corrupt latent data representations. Previous work has attempted to alleviate this problem by exploiting contrastive learning, the pair building of which is critical. However, existing methods either conduct sample-level processes and then use the resultant subset to construct pairs or directly perform pair-level selecting using a fixed threshold, both leading to sub-optimal pairing and subsequent representation learning. To address this issue, we propose a novel adaptive contrastive learning method (ACL) working at the pair level to select contrastive pairs adaptively. Specifically, we consider the model's learning status to adjust the confidence threshold in a self-adaptive manner instead of fixing it. Then, towards the ineffectiveness of the thresholding method on unconfident pairs, we automatically apply instance-specific temperature to boost the confidence of accurately-predicted samples and their pairs. We further introduce temporal cross-ensembling to handle the impact of noisy labels on model predictions. As a result, diverse pairs are correctly selected for contrastive learning to induce discriminative representations robust to various types of label noise. Extensive experimental results on several standard benchmarks and real-world datasets indicate the superiority of ACL, especially in extremely noisy scenarios.
Recently, image-to-image translation models have achieved great success in terms of content consistency and visual fidelity. However, in most of these tasks, the inaccuracy of sketches and the high cost of fine semantic masks acquisition limit the large-scale use of image translation models. Therefore, we propose to use graffiti that combines the advantages of sketches and semantic masks as model input. Graffiti reflects the general content of an image using lines and color distinctions, with some unlabeled regions. However, due to the large number of unknown areas in the graffiti, the generated results may be blurred, resulting in poor visual effects. To address these challenges, this paper proposes a multi-task framework that can predict unknown regions by learning semantic mask from graffiti, thereby improving the quality of generated real scene images. Furthermore, by introducing an edge activation module, which utilizes semantic and edge information to optimize the object boundaries of the generated images, the details of the generated images can be improved. Experiments on the Cityscapes dataset demonstrate that our multi-task framework achieves competitive performance on graffiti-based image generation task.
Contrastive Language-Image Pre-training (CLIP) starts to emerge in many computer vision tasks and has achieved promising performance. However, it remains underexplored whether CLIP can be generalized to 3D hand pose estimation, as bridging text prompts with pose-aware features presents significant challenges due to the discrete nature of joint positions in 3D space. In this paper, we make one of the first attempts to propose a novel 3D hand pose estimator from monocular images, dubbed as CLIP-Hand3D, which successfully bridges the gap between text prompts and irregular detailed pose distribution. In particular, the distribution order of hand joints in various 3D space directions is derived from pose labels, forming corresponding text prompts that are subsequently encoded into text representations. Simultaneously, 21 hand joints in the 3D space are retrieved, and their spatial distribution (in x, y, and z axes) is encoded to form pose-aware features. Subsequently, we maximize semantic consistency for a pair of pose-text features following a CLIP-based contrastive learning paradigm. Furthermore, a coarse-to-fine mesh regressor is designed, which is capable of effectively querying joint-aware cues from the feature pyramid. Extensive experiments on several public hand benchmarks show that the proposed model attains a significantly faster inference speed while achieving state-of-the-art performance compared to methods utilizing the similar scale backbone. Code is available at: https://github.com/ShaoXiang23/CLIP_Hand_Demo.
Despite significant progress in video question answering (VideoQA), existing methods fall short of questions that require causal/temporal reasoning across frames. This can be attributed to imprecise motion representations. We introduce Action Temporality Modeling (ATM) for temporality reasoning via three-fold uniqueness: (1) rethinking the optical flow and realizing that optical flow is effective in capturing the long horizon temporality reasoning; (2) training the visual-text embedding by contrastive learning in an action-centric manner, leading to better action representations in both vision and text modalities; and (3) preventing the model from answering the question given the shuffled video in the fine-tuning stage, to avoid spurious correlation between appearance and motion and hence ensure faithful temporality reasoning. In the experiments, we show that ATM outperforms existing approaches in terms of the accuracy on multiple VideoQAs and exhibits better true temporality reasoning ability.
The task of dynamic scene graph generation (DSGG) aims at constructing a set of frame-level scene graphs for the given video. It suffers from two kinds of spurious correlation problems. First, the spurious correlation between input object pair and predicate label is caused by the biased predicate sample distribution in dataset. Second, the spurious correlation between contextual information and predicate label arises from interference caused by background content in both the current frame and adjacent frames of the video sequence. To alleviate spurious correlations, our work is formulated into two sub-tasks: video-specific commonsense graph generation (VsCG) and causal inference (CI). VsCG module aims to alleviate the first correlation by integrating prior knowledge into prediction. Information of all the frames in current video is used to enhance the commonsense graph constructed from co-occurrence patterns of all training samples. Thus, the commonsense graph has been augmented with video-specific temporal dependencies. Then, a CI strategy with both intervention and counterfactual is used. The intervention component further eliminates the first correlation by forcing the model to consider all possible predicate categories fairly, while the counterfactual component resolves the second correlation by removing the bad effect from context. Comprehensive experiments on the Action Genome dataset show that the proposed method achieves state-of-the-art performance.
Spatio-temporal video grounding (STVG) aims to localize the spatio-temporal object tube in a video according to a given text query. Current approaches address the STVG task with end-to-end frameworks while suffering from heavy computational complexity and insufficient spatio-temporal interactions. To overcome these limitations, we propose a novel Semantic-Guided Feature Decomposition based Network (SGFDN). A semantic-guided mapping operation is proposed to decompose the 3D spatio-temporal feature into 2D motions and 1D object embedding without losing much object-related semantic information. Thus, the computational complexity in computationally expensive operations such as attention mechanisms can be effectively reduced by replacing the input spatio-temporal feature with the decomposed features. Furthermore, based on this decomposition strategy, a pyramid relevance filtering based attention is proposed to capture the cross-modal interactions at multiple spatio-temporal scales. In addition, a decomposition-based grounding head is proposed to locate the queried objects with less computational complexity. Extensive experiments on two widely-used STVG datasets (VidSTG and HC-STVG) demonstrate that our method enjoys state-of-the-art performance as well as less computational complexity. The code has been available at https://github.com/TJUMMG/SGFDN.
Out-of-distribution~(OOD) detection is the key to deploying models safely in the open world. For OOD detection, collecting sufficient in-distribution~(ID) labeled data is usually more time-consuming and costly than unlabeled data. When ID labeled data is limited, the previous OOD detection methods are no longer superior due to their high dependence on the amount of ID labeled data. Based on limited ID labeled data and sufficient unlabeled data, we define a new setting called Weakly-Supervised Out-of-Distribution Detection (WSOOD). To solve the new problem, we propose an effective method called Topological Structure Learning (TSL). Firstly, TSL uses a contrastive learning method to build the initial topological structure space for ID and OOD data. Secondly, TSL mines effective topological connections in the initial topological space. Finally, based on limited ID labeled data and mined topological connections, TSL reconstructs the topological structure in a new topological space to increase the separability of ID and OOD instances. Extensive studies on several representative datasets show that TSL remarkably outperforms the state-of-the-art, verifying the validity and robustness of our method in the new setting of WSOOD.
In scenarios with long-tailed distributions, the model's ability to identify tail classes is limited due to the under-representation of tail samples. Class rebalancing, information augmentation, and other techniques have been proposed to facilitate models to learn the potential distribution of tail classes. The disadvantage is that these methods generally pursue models with balanced class accuracy on the data manifold, while ignoring the ability of the model to resist interference. By constructing noisy data manifold, we found that the robustness of models trained on unbalanced data has a long-tail phenomenon. That is, even if the class accuracy is balanced on the data domain, it still has bias on the noisy data manifold. However, existing methods cannot effectively mitigate the above phenomenon, which makes the model vulnerable in long-tailed scenarios. In this work, we propose an Orthogonal Uncertainty Representation (hOUR) of feature embedding and an end-to-end training strategy to improve the long-tail phenomenon of model robustness. As a general enhancement tool, OUR has excellent compatibility with other methods and does not require additional data generation, ensuring fast and efficient training. Comprehensive evaluations on long-tailed datasets show that our method significantly improves the long-tail phenomenon of robustness, bringing consistent performance gains to other long-tailed learning methods.
One of the major obstacles that hinders the current scene graph generation (SGG) performance lies in the severe predicate annotation bias. Conventional solutions to this problem are mainly based on reweighting/resampling heuristics. Despite achieving some improvements on tail classes, these methods are prone to cause serious performance degradation of head predicates. In this paper, we propose to tackle this problem from a brand-new perspective of dark knowledge. In consideration of the unique nature of SGG that requires a large number of negative samples to be employed for predicate learning, we design to capitalize on the dark knowledge contained in negative samples for debiasing the predicate distribution. Along such vein, we propose a novel SGG method dubbed Dark Knowledge Balance Learning (DKBL). In DKBL, we first design a dark knowledge balancing loss, which helps the model learn to balance head and tail predicates while maintaining the overall performance. We further introduce a dark knowledge semantic enhancement module to better encode the semantics of predicates. DKBL is orthogonal to existing SGG methods and can be easily plugged into their training process for further improvement. Extensive experiments on VG dataset show that the proposed DKBL can consistently achieve well trade-off performance between head and tail predicates, which is significantly better than previous state-of-the-art methods. The code is available in https://github.com/chenzqing/DKBL.